Gamma Spectrum Analysis Based On Genetic Neural Network

Gamma spectrum analysis techniques have important applications inthe analysis and identification of nuclear material, safety testing of nuclear facilities,environmental radioactivity monitoring and prevention of nuclear terrorism.This thesis briefly reviewed radioactive detection and solid-state detector,gamma spectroscopy solution spectral method and BP neural network. To analysis andstudy the measured gamma spectroscopy, through using BP neural network whichoptimize by genetic algorithm.The main conclusions obtained are as follows:1) The optimal solution of BP neural network structure weights and thresholds isobtained and the adaptive ability, generalization ability of the network is improved byusing of genetic algorithm to optimize BP neural network. The neural networktoolbox of MATLAB is used to build the BP neural network structure which isoptimized by adaptive genetic algorithm. The best network parameter settings aremeasured: using three-layer BP network, having73hidden layer nodes, selectinglogsig or tansig as the hidden layer the connection weights, choosing logsig as outputlayer connection weights and trainrp as the training function. The best individual inthe population is retained through improving genetic algorithm fitness function andGenetic Operators. The accuracy of the network can be achieved10-3, and the numberof training steps can be reduced to40steps. Comparing with the non-optimized BPneural network, the adaptive genetic algorithm optimize BP neural network greatlyenhance the network running speed, avoid the network to falling into the localminimum and improve the adaptive ability and generalization ability of the network.2) Using adaptive genetic algorithm to optimize BP neural network gammaspectrum analysis and the full-spectrum input method for analysis, the accuracy of theresults is improved. In the recognition process, smooth the gamma spectrum, makespectrum stabilization and remove the bottom processing, reduce the error caused bystatistical fluctuations and environmental background. Use the Full-spectrum inputmethod for data input. Gamma spectroscopy each count as a neural network inputvalue, do not discard any data. The most important advantage is avoiding the gammaenergy spectrum peak search, energy calibration and efficiency calibration. Theaccuracy of the results of the analysis of the gamma spectrum of the neural network is greatly improved by eliminating the analytical error caused by searching peak, energyand efficiency calibration.3) The analytical accuracy of BP neural network gamma spectrum whichoptimized by adaptive genetic algorithm is better than the ordinary BP neural network.Eight radionuclide gamma energy spectrum (Am, Eu, Co, Cs, Ba, Ra, Th and U) andthe corresponding background spectrum are constituted the basis of the training datamatrix and formed the AGA-BP neural network. The affection of closeness of fullpeak, the radionuclide trained or not, statistical fluctuation, measuring distance andShield are taken in to consideration in the study of gamma spectral identification. TheAGA-BP neural network is able to identify all radionuclides under laboratoryconditions and the error of quantitative analysis is less than1%. Comparing with BPneural network solution spectrum results, finding analysis accuracy of AGA-BPneural network was better than ordinary BP neural network results.